Add the Slack focus board as the dashboard's Slack tab

CHANGELOG.md

@@ -9,6 +9,74 @@ The version is carried in `pyproject.toml` and `src/assistant/__init__.py`
 
 ## [Unreleased]
 
+### Added
+- Add a **Slack** tab to the dashboard, ported from the standalone focus-board
+  tool. It lists the Slack threads where the next move is yours, in Do first,
+  Promise, This week, and Optional, plus Done by you, Just read, Set aside, and
+  Auto-resolved.
+  - **Refresh from Slack** POSTs `/slack-focus/refresh` to `todo-server.py`,
+    which starts `bin/slack-focus.py` in the background under zsh with
+    `~/.zprofile` sourced, so `$SLACK_USER_TOKEN` and `gh` resolve under
+    launchd. The page polls `/slack-focus/status` and shows each step. A second
+    click while a refresh runs starts nothing.
+  - **Done** and **Undo** POST `/slack-focus/done` and `/slack-focus/undone`.
+    A checked-off row comes back when its thread gets a new reply or edit.
+  - The pipeline lives in `src/assistant/slack_focus/`. Collect and triage are
+    rules only; the judge is one model call over the surfaced candidates,
+    tools off, metered as `caller="slack_focus"`, with verdicts cached by
+    content. If the model fails, the raw rows show. It uses the balanced
+    model tier with one attempt; the focus-board ran your `claude` alias and
+    retried once. `bin/assistant-llm.py` routes it as the `slack_judge`
+    feature.
+  - The Slack client refuses any method outside a read allowlist before a
+    request leaves the machine, and a test checks that no module names a
+    common Slack write method.
+  - Everything about the user is in the new `slack_focus` block of
+    `~/.assistant/comms/config.json`: watched channels, bot names, skipped
+    channel names, GitHub repos and hosts, and Jira projects. The Slack ID,
+    name, and workspace come from the token. Jira project keys and GitHub
+    hosts used to be hard-coded.
+  - Files live in `~/.assistant/slack-focus/`.
+  - Each row stores its thread's latest message, so a check-off keeps its
+    key while a new pull renumbers threads. The key formula matches the
+    focus-board's, so its `done.json` carries over.
+  - A refresh counts as running only while it holds `refresh.lock`, so a
+    crash or a reused process ID can't leave the button stuck.
+  - Any Slack error other than a deleted or private thread or channel stops
+    the refresh with that error; `ratelimited` answers are retried. The
+    focus-board carried on with a partial pull. A refresh stops after 20
+    minutes.
+  - The Slack routes refuse `Origin: null`, which the other routes allow
+    for file:// pages.
+  - Times on the tab show in local time.
+  - Promises made in DMs read "DM" or "group DM" instead of a channel ID.
+  - Jira uses a bearer token unless `email_env` names the env var with your
+    account email; the focus-board always looked for `$JIRA_EMAIL`.
+  - A new reply in a DM thread that started before the window now shows up.
+    The focus-board dropped it.
+  - A message counts as answered only when you posted after it. The
+    focus-board also counted a message of yours from up to 60 seconds
+    before, which hid quick DM asks.
+  - A direct @-mention of you outranks an earlier group mention in the same
+    thread.
+  - DMs and threads are read to the end. The focus-board stopped at 200
+    messages.
+  - A model answer of `[]` sets every new row aside as noise. The
+    focus-board treated it as a failure and showed raw rows.
+  - A promise the model moves to Do first ranks below direct asks instead
+    of above every row.
+  - A refresh error never includes the Slack token.
+  - If a refresh launch hangs before it starts, the next **Refresh from
+    Slack** stops it and starts a new one.
+  - The unused low-signal hints are gone.
+  - A `config.json` that isn't valid JSON shows that on the tab, with the
+    line, instead of the setup steps next to the last list.
+  - "Last pulled" shows when Slack was last pulled, also after a
+    `--triage` run.
+- A row's text now quotes the other person's latest message, not yours. The
+  focus-board compared author names with the label "you", which never
+  matched a real name.
+
 ### Changed
 - Read agent sessions from cmux first, with mukul-memory's registry as a
   fallback. cmux 0.64.25 records agent sessions itself, and

README.md

@@ -105,6 +105,8 @@ The whole thing optimizes one number downward, computed mechanically each day: *
 
 **🧠 Turn corrections into rules, then into memory.** `lesson-extractor.py` scans your recent Claude Code transcripts and the action ledger for corrections, confirmations, and recurring questions, and distills lesson candidates into a proposals queue for you to review. Confirm one and it's routed to the right store (one of `claude` / `assistant` / `ffp` / `archffp` / `assistant-repo`), then mirrored into the Obsidian vault and synced to the cross-machine memory repo that feeds Mem0 semantic memory. Nothing is added without your confirmation.
 
+**📌 See which Slack threads need you.** The dashboard's **Slack** tab lists the threads where the next move is yours: direct asks, @-mentions, DMs, promises you made, and unread posts in channels you watch. Rules sort every thread for free; one bounded model call writes short titles and sets noise aside. Threads whose linked pull request merged or ticket closed drop to **Auto-resolved**. Select **Done** to check a row off; it comes back if the thread gets a new reply. It's read-only: it never posts, reacts, or marks anything read. See [Slack focus list](#slack-focus-list).
+
 **👋 Nudge stalled work and move the safe stuff forward.** Each pulse the Observer emits one verdict per workspace, and Python — not the model — turns it into an action. `ready_for_merge` queues `/merge-when-ready`; `ready_for_cleanup` sends `/cleanup` (only on workspaces *it* queued the merge for, and only once a work receipt exists); `stranded` nudges the idle agent with what failed and a retry; `needs_user` surfaces an awaiting card and does nothing else. Autonomy is fenced by a back-off list, the work-receipt gate, the assistant-merge ledger, and `NO_INGEST_GUARD`.
 
 ## Setup — connect your sources (opt-in)
@@ -168,6 +170,43 @@ Default ingestion is **@-mentions + DMs only**. For full channel/group message i
 
 </details>
 
+### Slack focus list
+
+The Slack tab needs a Slack **user** token, because only user tokens can search messages, and a `slack_focus` block that says what to watch. Nothing about you is hard-coded: your Slack ID, name, and workspace come from the token.
+
+1. Create or reuse a Slack app with these **user** token scopes: `search:read`, `channels:history`, `groups:history`, `im:history`, `mpim:history`, `channels:read`, `groups:read`, `im:read`, `mpim:read`, `users:read`, and `usergroups:read`. Install it and copy the `xoxp-` token.
+2. Add `export SLACK_USER_TOKEN=xoxp-…` to `~/.zprofile`. The Refresh button sources that file, so the token never lands in a config file or plist.
+3. Add a `slack_focus` block to `~/.assistant/comms/config.json`:
+
+```json
+{
+  "slack_focus": {
+    "days": 7,
+    "title": "Slack focus",
+    "watched_channels": {"C0123456789": "my-team"},
+    "bot_authors": ["Deploy Bot"],
+    "skip_channels": ["deployments"],
+    "github_repos": ["my-org/my-repo"],
+    "github_hosts": ["github.com"],
+    "jira": {"enabled": false, "base_url": "https://jira.example.com", "project_keys": ["ABC"],
+             "token_env": "JIRA_API_TOKEN"}
+  }
+}
+```
+
+| Setting | What it does |
+|---|---|
+| `days` | How far back to look, 1–30. Default 7. |
+| `title` | The heading inside the Slack tab. The tab itself is always labeled **Slack**. |
+| `watched_channels` | Channel ID → a label for your own reference. Every thread there counts, not only the ones that mention you. Mentions and DMs are always checked. |
+| `bot_authors` | Display names of bots. A thread whose latest message is from one of them is treated as a bot post. |
+| `skip_channels` | Channel-name fragments to treat as bot noise, such as `deployments`. |
+| `github_repos` | Repos where your merged PRs can retire an ask ("already done"). Uses `gh`. |
+| `github_hosts` | Hosts whose pull request links count, for GitHub Enterprise. Default `github.com`. |
+| `jira` | Optional. With `enabled` and `project_keys`, a thread whose linked tickets are all done is auto-resolved. The token comes from the env var named in `token_env` (default `JIRA_API_TOKEN`), sent as a bearer token. For Jira Cloud, set `email_env` to the env var that holds your account email, such as `JIRA_EMAIL`, to use basic auth. `base_url` must start with `https://`; it falls back to `$JIRA_BASE_URL`. |
+
+Open the **Slack** tab and select **Refresh from Slack**. A 7-day pull takes about 4 minutes, plus up to a minute for the model step; a refresh stops after 20 minutes. A Slack error, such as a missing scope or a rate limit that doesn't clear, stops the refresh and shows the error on the tab, so the list is never built from a partial pull. To refresh from a terminal, run `python3 bin/slack-focus.py` (`--triage` reuses the last pull, `--no-judge` skips the model). Files live in `~/.assistant/slack-focus/`. To route the model step to Droid, run `bin/assistant-llm.py set droid --feature slack_judge`. Moving from the standalone focus board? Copy its `out/done.json` to `~/.assistant/slack-focus/done.json` to keep your check-offs.
+
 ## Config knobs you own
 
 | Knob | Default | Where | What it does |
@@ -177,6 +216,7 @@ Default ingestion is **@-mentions + DMs only**. For full channel/group message i
 | Planner autoDispatch | **OFF** | `~/.assistant/comms/config.json` → `{"planner":{"autoDispatch":true}}` | OFF (safe default) = a stalled goal *stages a decision* for your review. ON = whitelisted steps dispatch unattended overnight. Flip it on only after watching the planner behave. |
 | JIRA scope | your own issues | `JIRA_JQL` env var | Widens the JQL beyond `assignee/reporter/watcher = you`. |
 | Slack channel ingest | mentions + DMs only | `SLACK_INGEST_CHANNELS=1` env (+ manifest edit) | Opts into full-channel message ingestion. |
+| Slack focus list | not set up | `~/.assistant/comms/config.json` → `slack_focus` block + `SLACK_USER_TOKEN` | Which channels, bots, repos, and Jira projects the Slack tab uses. See [Slack focus list](#slack-focus-list). |
 | Connector cadence / caps | 60s · 200 events · 10 pages | `~/.assistant/comms/config.json` → `connectors` block | Per-connector poll cadence and batch caps (module constants are only the fallback). |
 
 **Strategist (M6, the LLM drafter).** In `~/.assistant/comms/config.json` → `{"strategist": {…}}`: `enabled` (default `true`), `dailyCostCeilingUsd` (the day's LLM budget — the Strategist is *shed first* when it's hit, never the Observer), plus per-goal `budget.maxStrategistCallsPerDay` (default `1`) in `assistant-goals.json`. It ships enabled, but on the safe `autoDispatch`-off default it only upgrades *staged-decision* text you review — never unattended work.
@@ -187,6 +227,7 @@ Open the dashboard at **http://127.0.0.1:9876** (localhost only; served by `bin/
 
 - **Sessions tab** — your default view counts open cmux workspaces, not GitHub alerts. It groups workspaces into **Needs your answer**, **Assistant tasks**, **Ready to close**, and **Paused**. Each card shows one next action. Click **Go to session** to focus its workspace, or expand the card for details. The finish prompt highlights older pending work without interrupting productive sessions or automatically closing anything.
 - **Notifications tab** — a separate inbox for GitHub pull requests and other alerts. Its counts don't add to your session workload. Expand a group's history for its existing controls; receipts, the FYI digest, and health remain available here.
+- **Slack tab** — the threads where the next move is yours, in **Do first**, **Promise**, **This week**, and **Optional**. Select **Refresh from Slack** to pull again; the button shows each step. **Done** checks a row off until its thread gets a new reply.
 - **Connections tab** — every connector as **Connected** / **Available, not connected** / **Needs attention**, with the how-to-connect hint inline.
 - **`/goal` skill** — add / list / rerank / pause goals (`/goal add "…" --outcome "…"`). Goals feed the planner and boost decision ranking. Automation can only *propose* goal changes; only you (this skill) edit the store in place.
 - **One-tap accept** on a decision routes through the todo-server, executes the recommended action class (draft-only for any external send in this tree), and ledgers the transition. Reject / snooze do the same.
@@ -276,6 +317,7 @@ expanded original questions, and unchanged controls.
 - `bin/build-morning-brief.py` — rebuild today's brief on demand; `bin/plan-next-actions.py` — run the goals planner on demand
 - `bin/interrupt-gate.py` — the sole push chokepoint (consulted for every would-be notification)
 - `bin/todo-server.py` — the localhost dashboard + JSON API at http://127.0.0.1:9876
+- `bin/slack-focus.py` — refresh the Slack tab's list on demand (the dashboard's Refresh button runs it)
 - `bin/assistant-curator.py` — lesson/rule management across all five stores
 - `bin/tool-dispatch.py` — named tool dispatcher (`bin/tools-manifest.json`)
 - `install.sh --apply` — wires everything up; symlinks skills; writes/copies plists (never loads them)

bin/assistant-llm.py

@@ -23,6 +23,7 @@
     "strategist",
     "narrator",
     "lessons",
+    "slack_judge",
 )